iFlytek Unveils Spark X2.6, the Xingchen TaaS Platform and the SparkAudio 1.0 Speech Model

At the 1024 developer festival, which opened on October 22, 2026, at the National Convention Center in Beijing, iFlytek rolled out a package of four announcements. Company chairman Liu Qingfeng presented Spark X2.6 with a "device + cloud" architecture, the Xingchen TaaS platform, the SparkAudio 1.0 speech foundation model, and a multimodal interaction technology for embodied AI. Separately, at a meeting with the developer ecosystem, iFlytek's open platform showed a new generation of human-machine interaction tools.

The most telling item on the list is not another model version but Xingchen TaaS. The release slogan reads "make the token a true productive force for thousands of industries." It is a bid to change the unit in which the company talks to its customers.

The token as the unit of delivery

TaaS stands for Token as a Service. The logic is clear to anyone who has seen an API bill: a business does not need an abstract "large model", it needs a predictable price per unit of useful work. iFlytek is building an entire platform around the token rather than simply putting a price on an endpoint. Everything that sits between the model and a specific production scenario is gathered under the Xingchen name.

For the Chinese market, this is a logical move. Inference prices in China have collapsed over the past two years, and a model as a standalone product barely earns any margin. The money has to come from integration, reliability and industry-specific tooling. That is the layer iFlytek is trying to package as a service.

Spark X2.6: part of the work stays on the device

Spark X2.6 is billed as "端云一体", meaning a unified architecture on the device side and in the cloud. The idea is simple. Light requests are processed locally, where latency and privacy matter, while heavy ones go to the cloud. For a company whose history grew out of voice assistants, learning tablets, voice recorders and translators, this is a natural setup: iFlytek already has a fleet of its own hardware to run the model on.

The version number 2.6 suggests an incremental update to the Spark X line rather than a generational change. But the link to TaaS changes the context. Tokens generated on the device and in the cloud can be metered and sold within a single system.

Voice and body: SparkAudio and the embodied interface

On the speech front, the company released SparkAudio 1.0, a foundation model for audio. This is the territory where iFlytek has the longest track record: speech recognition and synthesis were its core product long before the LLM era. Now the same expertise is being packaged as a foundation model rather than a set of separate engines.

The third line concerns robots. The multimodal interaction technology for embodied AI is meant to let a machine perceive voice, images and context at the same time and respond to them. For humanoids and service robots, the conversational interface remains one of the weakest points, and iFlytek's voice work looks like a strong asset here.

Key items in the announcement:

• Spark X2.6: a version of the flagship model with a "device + cloud" architecture
• Xingchen TaaS: a platform where the token is the unit of delivery
• SparkAudio 1.0: the first version of the speech foundation model
• Multimodal interaction for embodied systems
• A new generation of human-machine interaction tools on the open platform

Who needs this, and when

Developers and integrators in China should view Xingchen TaaS as an attempt to offer a ready-made pipeline from model to industry application, not just another API. The main question for evaluating the platform is the final cost of a token in a real scenario, including the local part running on the device.

The hybrid approach in X2.6 will interest smart-device makers: it lets them keep some scenarios offline and avoid paying for the cloud on every utterance. Robotics teams should look at SparkAudio and the multimodal stack as a ready-made voice layer.

Outside China, the effect will be indirect. But the trend itself is visible: model providers are moving from selling intelligence to selling the token as a measurable unit of production, and iFlytek announced it on the biggest stage of its calendar.